Writing a good brief

Why it matters

The copilot is only as good as what you ask. A vague request gets a reasonable guess. A clear brief gets a design you can defend.

The five things a good brief says

1. The shell. Building size, clear height, column grid.
2. The storage target. How many positions, of what type, how high.
3. The flow. Where goods come in and leave. Which wall the docks are on.
4. The support spaces. Staging and dispatch floors, offices, charging, amenities.
5. The constraints. The forklift and aisle width, keep-outs, anything the site forces.

A brief that works

Design a distribution centre in a 120 × 100 m shed, 12 m clear, columns on a 12 × 10 m grid.

Storage: about 4 500 selective pallet positions, 4 levels, counterbalance trucks with 3.8 m aisles. A forward pick face of about 200 positions near dispatch.

Flow: three receiving doors on the south wall at the west end, six dispatch doors on the south wall at the east end. A receiving floor in front of the receiving doors and a dispatch floor in front of the dispatch doors.

Support: an office for 8 people near the receiving doors, a packing bench between the pick face and dispatch.

Check the design when you are done and fix anything major.

On a brief like this, the copilot plans the whole floor first (a block plan), then builds it block by block, then measures it.

Tips

• Give numbers. "About 4 500 positions" beats "lots of storage".
• Name the wall. "Docks on the south wall" removes a guess.
• Name things. "A zone called Reserve" lets you refer to it later: "make Reserve 5 m shorter".
• One idea per follow-up. "Now add a cross aisle every 8 bays in Reserve."
• Ask it to check. "Check the design" runs the same measurements it uses on itself.
• Ask what it assumed. "What did you assume?" gets a straight list.
• Use memory for habits. "Remember that this client always uses 3.8 m aisles."

Asking for a what-if

The copilot can run the whole comparison workflow in one message:

Clone this scenario as '+25 % demand', raise the clone's two inbound truck profiles and its Interstate order type by 25 %, run the clone for 7 days with 3 replications on seed 42 and racks 30 % full, then compare it with the baseline. Report the headline differences with their 95 % bands and say which are real.

In the demo, that took eight tool calls and under a minute. A scaled profile is noted on the profile itself ("scaled x1.25 by the copilot") so a second request does not scale it twice.

To compare again later without changing anything, say exactly that: "Compare the two existing runs. Change nothing, run nothing."

When it goes wrong

• It refused something. Read the quoted reason. It nearly always names the fix: a distance, a level count, a zone that is in the way.
• It ran out of steps. Click Continue.
• It is heading the wrong way. Type a correction while it runs, or Stop and rephrase.
• The result is wrong. Say what was wrong, specifically: "Reserve overlaps the office, shorten it by 6 m". The copilot's changes are not on the editor's undo list, so ask it to remove what it built, or delete it yourself. For a big experiment, clone the scenario first.

Words you just learned

Brief · block plan.